Error-correcting codes and applications to large scale classification systems

Abstract

In this thesis, we study the performance of distributed output coding (DOC) and error-Correcting output coding (ECOC) as potential methods for expanding the class of tractable machine-learning problems. Using distributed output coding, we were able to scale a neural-network-based algorithm to handle nearly 10,000 output classes. In particular, we built a prototype OCR engine for Devanagari and Korean texts based upon distributed output coding. We found that the resulting classifiers performed better than existing algorithms, while maintaining small size. Error-correction, however, was found to be ineffective at increasing the accuracy of the ensemble. For each language, we also tested the feasibility of automatically finding a good codebook. Unfortunately, the results in this direction were primarily negative.